Predicting drug release from polymeric long-acting injectables using a machine learning approach to decode formulation-performance relationships.

Polymeric long-acting injectables (LAIs) are among the most impactful yet least predictable drug delivery systems, as their performance emerges from a widely unresolved combination of polymer microstructure, manufacturing history, and assay conditions. To address this challenge, we introduce a hybrid machine learning framework that combines data-driven prediction with mechanistically relevant release theory to systematically decode LAI performance. Using a curated dataset of 1,537 literature-derived release profiles and 35 descriptors, our random forest model achieved robust performance with a mean (R² = 0.86 ± 0.16). Interpretable SHAP analysis confirmed a critical insight: the release apparatus can shape measured kinetics as strongly as the formulation itself, highlighting a major hurdle in current standardization. The framework's prospective utility was examined through validation using the commercial product Risperdal Consta®, resolving method‑dependent shifts and achieving the desired outcome. Stress/exploratory testing of newly in-house prepared donepezil-loaded PLGA microspheres (intermediate release) yielded high prediction accuracy (MAE = 0.06), whereas prediction accuracy was notably challenged for slow-release formulations. Further investigation revealed that incorporating only three early-stage measurements reduced prediction error for slow-release formulations, providing a practical route for error control and offering a perspective on the use of literature-derived training models for ongoing formulation prototyping. These findings position hybrid artificial intelligence as a promising strategy for the rational development of polymeric LAIs, bridging polymer science and drug release performance while replacing empirical trial-and-error with data-driven precision.